🤖 ComfyAgent AI Assistant

First, stop looking for wires, because there are none. The node called 🤖 ComfyAgent AI Assistant has zero inputs and zero outputs - its info_schema is empty and its Python class is literally a noop that returns nothing. The source is honest about it: the class is named ComfyAgentExtensionNode, and its entire job is to register the extension so ComfyUI loads it. You never drag this onto a graph. The product is the floating blue "AI Chat" button that appears in the top-right of the ComfyUI web app once the pack is installed, plus the backend routes behind it.

That's an unusual pattern and it's worth being clear-eyed about before you install. This isn't a prompt enhancer or a captioner that sits in your workflow like the LLM-in-the-graph tools you may have seen. It's an agent: a chat sidebar that can reach into your running ComfyUI, read your canvas, and do things to it. What it actually does, per the README and the code: introspect your workflow node-by-node, look up what any node class does, let you paste a node or a subgraph into the chat, search the custom-node registry, clone a GitHub repo and have ComfyUI Manager auto-install it (with your approval), back up and try to fix a broken node, and restart the backend. The selling point is that the model talks about your graph, not a generic one - it can verify that a workflow has a model, positive/negative conditioning, a latent, sampler settings, a VAE and an output node before it claims your image will render.

How it works is where this diverges from most local-LLM ComfyUI projects. There is no local model, no GGUF download, no VRAM budget - the README and code show it's an API agent. It talks to OpenRouter (https://openrouter.ai/api/v1) by default with a model setting of auto, which queries OpenRouter's list of free models and picks the best available one (the priority list leans on Gemini and Llama :free variants). You can point it at any OpenAI-compatible custom endpoint and model instead. It uses function/tool calling, and here's the part I like: the agent has a mandatory checkpoint tool for multi-step work. After each step it has to review what's done, what's left, what connections matter, and then choose continue, ask_user, or stop. It can't silently keep working - if it chooses ask_user it hands control back to you. Actions that touch your credentials (api_key, a Civitai key, a GitHub token) need an explicit per-action grant in the UI, which is a decent safety instinct in a category of node that has historically shipped malware.

The inputs/outputs section of this article is one sentence: there are none. Everything lives in the settings panel and the slash commands - /model, /settings, /help, /session, /new. Get used to those; they're the only knobs you have.

Install is genuinely light. Via ComfyUI Manager, search ComfyAgent and install; or:

cd ComfyUI/custom_nodes
git clone https://github.com/NiklasNK-Creator/ComfyAgent

Then restart ComfyUI and hard-refresh the browser tab - the whole assistant is frontend JS, so an old cached tab won't grow the AI Chat button. No model files, and the only declared Python dependencies (aiohttp, Pillow) already ship with ComfyUI. What you do need is an OpenRouter API key, set through /settings or the panel.

The caveats are real, so read them. First, OpenRouter's free tier has a tiny credit balance, and the code is blunt about it: it caps max_tokens at 800 specifically because the default 65536 makes free accounts hit a 402. You will occasionally see a rate-limit error telling you to wait 60–120 seconds or switch models via /model. Second, key storage is only encrypted if the optional cryptography package is installed; otherwise settings fall back to a lightly obfuscated encoding - this is a brand-new, near-unknown pack (zero search impressions at the time of writing) that holds API keys and can restart your backend, so treat it like any fresh node: skim what it does and don't hand it more credentials than it needs. Finally, it's not open source in the usual sense - the license permits personal private use but forbids redistribution, so don't wrap it in your own pack.

Bottom line: it's a promising, very young take on the "assistant that operates ComfyUI itself" idea, and because it's API-based it costs you nothing in VRAM. If you want an LLM that writes prompts offline and uncensored, this is the wrong tool - that's the local-model job. If you want a chatbot that can look at your actual workflow, fix a node, and install what's missing, it's worth a weekend try.